Regional Climate Change Over Eastern Amazonia Caused by Pasture and Soybean Cropland Expansion
SOYBEAN CROPLAND MAPPING USING MULTI-TEMPORAL SENTINEL-1 DATA€¦ · SOYBEAN CROPLAND MAPPING...
Transcript of SOYBEAN CROPLAND MAPPING USING MULTI-TEMPORAL SENTINEL-1 DATA€¦ · SOYBEAN CROPLAND MAPPING...
SOYBEAN CROPLAND MAPPING USING MULTI-TEMPORAL SENTINEL-1 DATA
Mamta Kumari 1,*, C.S. Murthy 1, Varun Pandey 1, G.D. Bairagi 2
1 Agricultural Sciences & Applications Group, RS & GIS – Applications Area, National Remote Sensing Centre, Indian Space Research Organization, Hyderabad, India
2 Madhya Pradesh Council of Science and Technology, Bhopal, Madhya Pradesh, India
Commission III, WG III/10
KEY WARDS: Soybean, Sentinel-1, SAR, Classification, Phenology, SVM
ABSTRACT:
Soybean, a high value oilseed crop, is predominantly grown in the rainfed agro-ecosystem of central and peninsular India. Accurate
and up-to-date assessment of the spatial distribution of soybean cultivated area is a key information requirement of all stakeholders
including policy makers, soybean farmers and consumers. A methodology for timely assessment with high precision of soybean crop
using satellite data is yet not operational in India. In this scenario, synthetic aperture radar (SAR) has been shown to be a reliable
form of gathering crop information, especially during monsoon season. In this work, repeat coverage from the C-band Sentinel-1
satellite over Ujjain district, Madhya Pradesh is used for in-season soybean crop mapping along with other agricultural land-cover
types. The data were processed through four steps: (a) data preprocessing, (b) constructing smooth time series backscatter data, (c)
soybean crop classification using knowledge-based decision rule classifier and support vector machines (SVM) and (d) accuracy
assessment. The results indicated that the smooth VH backscatter profiles reflected the temporal characteristics of soybean crop
growing in the study region. Phenological characteristics were also derived from the smoothed S-1 VH backscatter time series to
segregate early and late sown soybean. This information was used as an input to a decision-rule classifier and SVM in order to
classify the input data into soybean and other crops. An overall accuracy of more than 80% using SVM and 75% using rule based
classifier, in Ujjain district was achieved. These results demonstrate the scope for using time-series S-1 VH data to develop an
operational soybean crop-monitoring framework.
1. INTRODUCTION
Soybean (Glycine max), being a high value oilseed, is also
known as the “Golden Bean” of the twentieth century. In the
recent past, soybean cultivation has increased manifold as
compared to any other oilseed crop in India and stands next
only to groundnut. Soybean is primarily grown in the rainfed
agro-ecosystem of central and peninsular India. About 53 per
cent of the cropped area falls under this crop in Madhya
Pradesh. In recent years, recurrent aberrant rainfall pattern and
intermittent dry spell, both in time and space, were observed in
these parts of country. These discrepancies inhibited the
successful cultivation of soybean crop and leads to significant
annual variation in cultivated area of soybean. Access to timely
and reliable data on crop distribution and its condition assumes
importance for enabling decision-makers to make informed
decisions with respect to imports, exports, and other policy
matters. Since, soybean is a short duration kharif season crops,
mapping of soybean crops and their growth assessment using
optical data is difficult due to consistent clouds. Mapping and
monitoring of soybean crop is important due to various reasons
such as, this crop is not covered in FASAL project; there is
requirement from the Ministry of Agriculture to increase the
number of crops for crop inventory; economic values of
soybean crop are very high; and soybean crop is highly relevant
from the crop insurance perspective.
Optical remote sensing is one of the most commonly used data
source for deriving crop information such as crop type map,
biophysical parameters, crop growth stages, yield etc. Crop
leaves are very sensitive to visible and infrared regions.
However, persistent cloud cover during the monsoon season
makes the availability of optical data difficult. This is almost
impossible to obtain a cloud-free optical image during this
period in a subtropical country like India. In this scenario,
Synthetic Aperture Radar (SAR) is the only option to map the
kharif season crops. Contrary to optical sensors, SAR sensors
are capable to acquire images under all weather conditions
which make it appropriate for crop mapping and monitoring
during rainy season (Liao et al., 2018). However, the use of
SAR data has not been established for agricultural applications,
as compared to optical data. There are several reasons, such as,
the complexity, cost and availability of SAR data, in addition to
the difficulty of data interpretation. The recently launched
Sentinel-1 with the advantage of free distribution provides more
opportunities to derive crop type map. Being the high spatial
and temporal resolution data, it allows not only to discriminate
different crop types but also to assess their growth status.
Previous studies performed the classification on Sentinel-1 data
for land cover mapping by utilizing backscatter values of single
dual polarimetric data (Abdikan et al. 2016, Whelen et al.,
2017), especially to map rice crop (Son et al., 2018, Nguyen et
al., 2017) and yielded maps with descent accuracy. However,
no studies have been carried out to map soybean crop using
Sentinel-1 data and very few studies are done using other SAR
data (McNairn et al., 2014). In this study, potential of dual
polarized Sentinel-1 backscatter data for in-season mapping of
soybean crop is investigated. We analyzed multi-temporal C-
band VV and VH polarized Sentinel-1 SAR data to evaluate the
backscatter behavior of soybean crop and to map its spatial
distribution in Ujjain district, Madhya Pradesh.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-3/W6, 2019 ISPRS-GEOGLAM-ISRS Joint Int. Workshop on “Earth Observations for Agricultural Monitoring”, 18–20 February 2019, New Delhi, India
This contribution has been peer-reviewed. https://doi.org/10.5194/isprs-archives-XLII-3-W6-109-2019 | © Authors 2019. CC BY 4.0 License.
109
2. METHODOLOGY
2.1 Study area
Ujjain district is situated in the heart of Malwa Plateau, Madhya
Pradesh at a general elevation of 527 meter above mean sea
level. The district area extends between the parallels of latitude
220 50' and 230 46' North and between the meridians of
longitude 750 08' and 760 16' East. The normal annual rainfall of
Ujjain district is 914.5 mm. Ujjain district receive maximum
rainfall (about 92.1%) during southwest monsoon period i.e.
June to November. The total geographical area of the district is
6,130.23 Sq. Km and divided in seven tehsil and six blocks. The
irrigation facilities in Ujjain district are moderate and
groundwater is the main source of irrigation in the district.
Ujjain district is mainly agriculture-based district and its
cropping pattern is diversified. Black cotton soils with heavy to
light texture are found in the whole area. Soybean is the most
dominant kharif season crop. Other crops grown in kharif
season are urad, maize, vegetables etc.
Fig 1. Study area: Ujjain district
2.2 Data collection
The input datasets for the analysis are listed hereunder;
2.2.1. Sentinel 1A data: This study used dual-polarized
(VV/VH) C-band imagery from the European Space Agency’s
(ESA) Sentinel-1A, covering 2017 soybean crop growing
season. The Sentinel-1A operates at a radar wavelength of about
5.6 cm (corresponding to frequency 5.405 GHz; C-band), with a
repeat coverage of 12 days. It has four imaging modes:
interferometric wide swath (IW), extra wide swath (EW), strip
map (SM), and wave (WV). This study used the
S1A_IW_GRDH product, which contains VH and VV
polarizations with a swath width of approximately 250 km and a
spatial resolution of 10 m; which provides backscatter
normalized with respect to area. Eleven dates (22 nos. images)
with a descending orbit during June to October 2017 were used
in the analysis. The incidence angle of the VH and VV images
acquired over the study region ranges from approximately 37–
45°. The details of the data collection dates are:01st June, 13th
June, 25th June, 19th July, 31st Jul, 12th Aug, 24th Aug, 05th Sep,
17th Sep, 29th Sep, 23rd Oct.
2.2.2 Optical data and Ancillary data: For the study area,
multi-temporal Resourcesat-2 AWiFS15-days NDVI composite
and Landsat-8 OLI data for the year 2017 (June to October)
were used to produce reference classification maps for
validating the Sentinel-1A soybean crop map. The Resourcesat
2 AWiFS15-days NDVI composite data of 56 m resolution was
obtained from NRSC. The Landsat 8 OLI data of 30 m
resolution was downloaded from the USGS website.
2.2.3 Ground reference data: Field information on the
distribution of different land cover classes in general and
soybean crop in particular, has been collected. A mobile
application - FASAL app –developed by NRSC was used
during field visit during September 2017. Various information
such as crop type, crop sowing and harvesting dates, field
moisture status, plant density etc., were collected during field
visit, using mobile app. This aided in faster and more efficient
collection of ground information about the crop along with field
photographs, as well as in building up a geodatabase with which
could be directly uploaded on to Bhuvan geoportal.
2.3 Satellite data processing
SAR derived time-series backscatter data are the key datasets in
this project as per the methodology framework depicted in
Figure 2.
2.3.1 Data pre-processing: The Sentinel-1A data were pre-
processed using ESA’s Sentinel’s Application Platform
(SNAP). Firstly, radiometric calibration was performed to
convert digital pixel values of VV and VH amplitude into sigma
naught (σ°) values that can be directly related to the radar
backscatter of the scene. Radiometric calibration is required to
remove the errors due to fluctuations in the transmitted power,
receiver gains, system noise and the illumination pattern of the
antenna. Speckles from the σ° data were filtered using adaptive
Lee filter with 5 × 5 moving window size to make the
interpretation of features easy. The terrain correction was
applied to compensate the distortions due to the side-looking
geometry of images. In this study, the Range Doppler Terrain
correction algorithm that uses the elevation data from the 1 arc-
second DEM product from the Shuttle Radar Topography
Mission (SRTM) was used for geocoding and terrain correction.
In this process, data are resampled and reprojected to the
Universal Transverse Mercator (UTM) coordinate system (zone
43 N) with a grid of 10 m spacing maintaining the 20 m spatial
resolution. The data were then mosaicked and subset over the
study region. Because Sentinel-1 data are not fully polarimetric,
polarimetric decomposition methods were not available to be
included in the analysis. Both polarizations were examined
separately as well as in combination.
Fig 2. Framework of methodology showing steps for soybean
crop classification in study area
2.3.2 Time series filtering to construct smooth time
series backscatter data: The backscatter time-series data
Data collection
• Multi-temporal Sentinel-1A
data (June to September,
2017)
• Single date Landsat-8 OLI
data (Mosaic of 20th-29th
September, 2017)
• Ground data (2017) using
mobile app
Data Pre-processing
• Radiometric correction
• Speckle noise filtering
• Terrain correction and projection
• Image mosaicking and layer stacking
Constructing smooth time series VH
data
• Temporal linear interpolation
• Noise filtering with wavelet transform
Soybean crop classification
• Training data selection/Rule formation
• Support Vector machine/ Rule based
classification
Accuracy assessment
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-3/W6, 2019 ISPRS-GEOGLAM-ISRS Joint Int. Workshop on “Earth Observations for Agricultural Monitoring”, 18–20 February 2019, New Delhi, India
This contribution has been peer-reviewed. https://doi.org/10.5194/isprs-archives-XLII-3-W6-109-2019 | © Authors 2019. CC BY 4.0 License.
110
possess the short-term influence of environmental conditions
and inherent noises in the Sentinel-1A data (such as noise due to
speckle). Filtering is required to reduce these noises that can
potentially affect the classification results. In current study,
wavelet transform, an adaptive noise reduction method, is used
to cope up the problem. It has certain advantages such as to
preserve the full non-linearity and non-stationarity in physical
space, which is more conducive to data denoising and
information extraction. It has been extensively used for the
denoising the time series of satellite data (Sakamoto et al. 2005;
Galford et al. 2008; Martinez & Gilabert 2009; Chen, Huang et
al. 2011; Chen, Son et al. 2011). The wavelet transform W(s, τ)
of a signal x (t) can be expressed as follows:
𝑊(𝑠, 𝜏) = ∫𝑥(𝑡)𝜓 (𝑡 − 𝜏
𝑠)𝑑(𝑡),
CZ: Calibration zones, SP: Seasonality parameters, AS:
Adaptation strength
Where s > 0 and τ ϵ R; x (t) is the analyzed input signal; ψ (t) is
the mother wavelet; and sand 𝜏are scaling and translation
parameters. A mother wavelet, Coiflet function (with order 4)
was preferred over Daubechies and Symlet functions for signal
decomposition, to get time profile of backscatter data for
deriving the crop phenological information (Sakamoto et al.
2005). The processed output is a smoothed backscatter signal,
which will be used for the extraction of different phenological
stages of soybean (e.g., a start of the season, heading time, and
length of season); in order to discriminate the soybean pixels
from the other land cover classes.
2.3.3 Extraction of Vegetation Phenology Parameters:
The three phenological indicators, viz., date of beginning of
season (DoS), date of maximum backscatter (DoM), and Length
of the season (LoS) are derived in this step. This allows
delineating the areas as “early planted soybean and late planted
soybean”, and all other areas were placed in a generalized “non-
soybean” class. These parameters are summarized in Table 1.
Table1. Phenological parameters for the rule-based
classification
2.3.4 Training data selection: This study focused on
delineating soybean crops (early and late planted) in Ujjain
district. The non-cropped areas, including orchards, other crops
(maize, vegetables), forests, built-up areas and water bodies,
were masked out using the smooth VH backscatter data to limit
the analysis to the cropped areas. The Support vector machine
(SVM) used training samples extracted from the filtered time
series VH backscatter data. The fuzzy C-mean clustering
algorithm (Dunn 1973; Bezdek 1981) was applied to classify
the extracted pixels into several clusters. The clusters reflecting
the pattern of the soybean crop being analyzed were selected to
train SVM for classification.
2.3.5 Support vector machines: Support Vector Machines
(SVMs), a non-parametric statistical learning method, has
recently been used in image classification for various
applications. Being a supervised learning method, the SVM
need user-defined parameters and each parameter has different
impact on kernels hence the classification accuracy of SVMs is
based upon the choice of the parameters and kernels. The
algorithm projects training data of the input space into a high-
dimensional space using a kernel function in which the classes
can be separable.
The basic idea behind SVMs is to separate the two different
classes through a hyperplane which is specified by its normal
vector ‘w’ determining the hyperplane orientation and the bias
‘b’. The optimal hyperplane can be given as⟨𝑤,⃗⃗⃗⃗ 𝑥 ⟩ + 𝑏 = 0 and
the corresponding decision function is𝑓(𝑥 ) = 𝑠𝑔𝑛(⟨�⃗⃗� , 𝑥⃗⃗ ⃗⟩ + 𝑏),
where x is a point lying on the hyperplane. The sign of f(x)
depends on the side of the hyperplane where the sample lies.
The norm of the normal vector was equal to the inverse of the
distance, of the closest sample of both classes to the hyperplane.
Hence, the optimal separating hyperplane with maximal margin
can be formulated as the quadratic optimization problem (min τ
(�⃗⃗� ) =1
2|�⃗⃗� |2) and solved using Karush Kuhn Tucker conditions
and Lagrange multipliers (Arfken 1985; Sundaram 1996).
For a non-linear case, a kernel function 𝐾(𝑥 , �⃗⃗� ) is introduced
to enable an efficient computation. The ‘kernel trick’ can be
applied since all feature vectors only occurred in dot products.
In current study, radial basis function (RBF) was used for SVM
classification of the VH backscatter data as it has demonstrated
to produce more accurate results than other functions such as
linear and polynomial kernels (Camps-Valls and Bruzzone
2005) for classification of remotely sensed data (Huang et al.
2002; Keerthi and Lin 2003; Pal and Mather 2005). The
parameters such as cost parameter (C), gamma (γ), bias term (r)
were used in SVM classification. The optimal values of these
parameters were identified to increase the classification
accuracy, by various permutations and combinations of these
parameters in different models for sensitivity analysis of SVM
architecture.
2.3.6 Accuracy Assessment: For validation and evaluation
of classification results, standard accuracy assessment measures
i.e., kappa coefficient, overall accuracy, omission error, and
commission error were used.
3. RESULTS & DISCUSSIONS
The current study attempts to delineate kharif soybean crop area
using SAR data. The 12-day time series VH backscatter from
Sentinel-1A SAR has been effectively used to delineate the
soybean crop area from other land use land cover types, in view
of soybean being short duration crop and grown in the during
monsoon season. The present study was carried out in Ujjain
district. Madhya Pradesh. The results obtained in this study are
described in this section.
3.1 Temporal soybean backscatter signature from
Sentinel-1 SAR Data
Figure 3 shows the colour composites, which are created by
using the multi-temporal SAR acquisitions in order to highlight
the temporal characteristics within the soybean fields. The red,
blue, and green coolers in the figure# correspond to the images
acquired during the sowing (June/July), Flowering/Pod
formation (August/September) and post harvested (October)
period, respectively.
Parameters Definition and Explanation
DoS During the crop growing season, the date of the beginning of
season is defined as the first localminima in time-series of
smooth VH backscatter data
DoM During the growing season, the date when backscatter reaches
a maximum value Isdefined as the local maxima in time-series
of smooth VH backscatter data.
This date must come after the date of the beginning of season,
where it reaches its local minimum backscatter value.
LoS The length of the season is defined as the number of days
difference between DoM and DoS.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-3/W6, 2019 ISPRS-GEOGLAM-ISRS Joint Int. Workshop on “Earth Observations for Agricultural Monitoring”, 18–20 February 2019, New Delhi, India
This contribution has been peer-reviewed. https://doi.org/10.5194/isprs-archives-XLII-3-W6-109-2019 | © Authors 2019. CC BY 4.0 License.
111
Fig 3. False colour composite of temporal SAR image
The profiles produced from the wavelet transform of VH
backscatter data revealed temporal discriminations between two
types of soybean crop (early sown and late sown) and those of
other land use cover types (e.g. orchards, water body and built-
up area). The VH backscatter intensities of soybean crops
showed a gradual increase of backscatter values after sowing on
reaching the maximum value during the pod formation of crops.
The backscatter intensities of soybean crops during the sowing
period were approximately −21 dB and those during the pod
formation periods (i.e. approximately 50 days after sowing)
were −15 dB, respectively. After sowing period, soybean crop
height increase with the unfolding of trifoliate leaves till the
start of flowering stage. Thus, the backscatter intensity showed
a significant increase during this vegetative phase due to the
greater the contribution of VH from volume scattering of
soybean plant’s internal canopies. After pod formation period,
the soybean plant was characterized by a decrease in the number
of leaves (i.e. ripening stages), the backscatter intensities
slightly decreased until the harvesting period. The VH
backscatter intensities from orchards and built-up areas were
relatively high. The backscatter values of water were also
distinguishable from those of soybean crop and other land use
cover types, showing very low VH backscatter values
throughout the year. Only VH backscatter temporal profile
shown consistent temporal pattern and shown an increase in
backscatter with crop growth, however, VV backscatter and
derivatives of backscatter reflected the inconsistent and abrupt
changes with crop growth in temporal backscatter profile.
3.2 Characteristics of smooth VH backscatter LUC
profiles
Fig 4 illustrated the some noise in raw VH, VV, VH/VV ratio
and VH+VV sum backscatter profiles in the signature of
soybean crop growth profile. Among all these backscatter
profiles, VH backscatter profile showed maximum potential by
resembling the growth profile of soybean. Therefore, only VH
backscatter profile was chosen for further analysis; to reduce the
computation time and complexity. The smooth curves (for VH
backscatter) obtained from the wavelet transform indicated that
the noise was significantly mitigated, as shown in Fig 5. The
cross-polarized VH backscatter was more sensitive to the
growth of soybean crop than the like-polarized VV backscatter,
characterizing the temporal variations of the soybean crop
during the season. The more sensitivity of the soybean in the
cross-polarized VH backscatter was mainly attributed to the
depolarization of signal that occurred within the soybean
canopy during the multiple reflection of the incoming signal
inside the vegetation volume.
.
Fig 4. Temporal backscatter of Sentinel-1A data (a) VH (b) VV
(c) VH/VV ratio and (d) VH+VV sum
Fig 5. Original and Smoothed temporal profile of backscatter
from soybean crop from different places
3.3 Thresholds Selection for rule based classification
The temporal backscatter signatures is shown in Figure 6 served
to define specific thresholds which were applied to the Sentinel-
1 time series in order to generate soybean crop area maps for the
growing season of 2017.
To map the different type of soybean crop areas based on
sowing time, the first two key parameters, DoS- backscatter at
the beginning of season and DoM- the peak (maximum) of VH
backscatter are very important. The amplitude backscatter
difference (Δσ; the difference between DoS and DoM) also
plays a significant role to decide the threshold. The
corresponding positions (date) and backscatter values of DoS
showed in the Figure 8.0. The peak and valley of VH
backscatter (i.e., DoS, DoM) within the one soybean-growing
cycle can be identified by local maxima. To eliminate
unrealistic peaks, a threshold for VH backscatter is required.
These two parameter thresholds (i.e., DoM ≥ 80 and Δσ ≥ 30)
were selected based on a conservative deduction from training
data. The third parameter, length of the season (LoS) which is
defined as the temporal distance(number of days difference)
between the date of beginning of season and the date when
maximum backscatter value is recorded. This temporal distance
has to be greater than the shortest possible soybean growing
cycle and smaller than the longest possible soybean growing
(a)
(d)
-13
-12
-11
-10
-9
-8
-7
-6
-5
-4
VH
+V
V S
um
back
scatt
er (
dB
)
-22
-20
-18
-16
-14
-12
01-J
un
13-J
un
25-J
un
07-J
ul
19-J
ul
31-J
ul
12-A
ug
24-A
ug
05-S
ep
17-S
ep
29-S
ep
11-O
ct
23-O
ct
VH
ba
ck
sca
tter (
dB
)
-13
-12
-11
-10
-9
-8
-7
-6
-5
-4
01-J
un
13-J
un
25-J
un
07-J
ul
19-J
ul
31-J
ul
12-A
ug
24-A
ug
05-S
ep
17-S
ep
29-S
ep
11-O
ct
23-O
ct
VV
ba
cksc
att
er (
dB
)
(b)
-10
-9
-8
-7
-6
-5
-4
-3
01-J
un
13-J
un
25-J
un
07-J
ul
19-J
ul
31-J
ul
12-A
ug
24-A
ug
05-S
ep
17-S
ep
29-S
ep
11-O
ct
23-O
ct
VH
/VV
ba
ck
sca
tter
ra
tio
(d
B)
(c)
0
20
40
60
80
100
120S
cale
d V
H b
acksc
atte
r
Original backscatter
smoothed backscatter
0
20
40
60
80
100
120
Sca
led
VH
bac
ksc
atte
r
Original backscatter
smoothed backscatter
0
20
40
60
80
100
120
Sca
led
VH
bac
ksc
atte
r
Original backscatter
smoothed backscatter
0
20
40
60
80
100
120
Sca
led
VH
bac
ksc
atte
rOriginal backscatter
smoothed backscatter
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-3/W6, 2019 ISPRS-GEOGLAM-ISRS Joint Int. Workshop on “Earth Observations for Agricultural Monitoring”, 18–20 February 2019, New Delhi, India
This contribution has been peer-reviewed. https://doi.org/10.5194/isprs-archives-XLII-3-W6-109-2019 | © Authors 2019. CC BY 4.0 License.
112
cycle. From crop calendar and expert knowledge about soybean
growing season in the study areas, the threshold of temporal
distance which is about 70 and 110 days, respectively. If all
three conditions are met then the pixel is classified as any type
of soybean, otherwise as non-soybean area. To discriminate
within soybean crop types, DoS and harvesting time window
was identified to decide whether it is early sown or late sown
soybean crop.
Fig 6. Smoothed temporal VH backscatter profile from late and
early sown soybean
3.4 Comparison of Sentinel-1A derived spatial
distribution of soybean crop and classification accuracies
The mapping results obtained from the classification of the
smooth time series Sentinel-1A VH backscatter data using rule
based and SVM indicated comparable spatial distributions of
soybean crop (both early and late planted) with those from the
ground reference data. In study area, soybean crop is extensively
grown in kharif season and occupied most of the study region,
with considerable variability in sowing time. The classification
maps produced from rule based and SVM (Fig 7) were
evaluated with the ground reference data (200 points) to assess
the accuracy of the mapping methods. The optimal values of
SVM parameters were identified (Kernel= RBF, γ =0.2, C =100,
Pyramid level=0.2, Classification probability = 0.5) for soybean
crop classification in study area. The results indicated that the
mapping results obtained from SVM was slightly better that
those from rule based. The overall accuracy and Kappa
coefficient achieved from SVM were respectively 83.1% and
0.74, while the values from rule based were 76.3% and 0.68,
respectively (Table 2). There were some error sources
exaggerating the mapping results, possibly attributed to the
small variation of incidence angle, ground reference error and
mixed-pixel issues. We could see that soybean fields, relatively
small and scattered, were mixed with other land use cover types,
such as other crops (urad) having same seasonal pattern,
orchards, canals and rivers in settlement areas. The mixed-pixel
and boundary effects could complicate the radar backscattering
coefficients, causing temporal uncertainty in discriminating
between signatures of soybean crops and other land-use types,
consequently leading to misclassification of soybean crop and
exaggerating mapping errors in these areas. Moreover, the 12-
day temporal resolution of Sentinel-1A is relatively long to
capture phenological information of soybean crop (e.g. sowing
and maximum growth periods), which might lead to an
increased uncertainty in soybean crop type classification,
especially areas of high variability in sowing. These results were
reaffirmed by the MP government’s soybean crop area statistics
of Ujjain district with the relative error in area values of 16%
and 8.5% for rule based and SVM, respectively.
Fig 7. Crop map showing spatial distribution of soybean along
with other land use cover type derived by (a) SVM (b) Rule
based classifier
Table 2. Accuracy assessment
4. CONCLUSIONS
The current study investigated the feasibility of multi-temporal
Sentinel-1A VH backscatter data for soybean crop classification
in Ujjain district, Madhya Pradesh. The results indicated that the
smooth VH backscatter profiles (wavelet transform-based)
could represent the temporal variations in growth profile of
soybean crop in the study region. Moreover, the smooth VH
backscatter time series could also segregate the early and late
sown soybean. As the mixed-pixel, sowing time variability and
crop with similar phenology could lower the level of
classification accuracy, this study demonstrated that rule based
and SVM classifier confirmed the validity for mapping soybean
in the region. The overall accuracy and Kappa coefficient
achieved from rule were, respectively, 76.3% and 0.68, while
the values from SVM were 83.1% and 0.74, respectively. These
results demonstrate the scope for using time-series Sentinel-1
VH data to develop an operational large-scale soybean crop
mapping framework.
0
20
40
60
80
100
120
Scale
d V
H b
ackscatt
er
Late Soybean
0
20
40
60
80
100
120
Scale
d V
H b
ackscatt
er
Early Soybean
0
20
40
60
80
100
120
Scale
d V
H b
ackscatt
er
Late Soybean
0
20
40
60
80
100
120
Scale
d V
H b
ackscatt
er
Early Soybean(a)
(b)
Year Soybean area (MP
state agriculture
portal) (lakh ha)
Sentinel-1 derived soybean area (lakh ha)
by SVM
classifier
by Rule based
classifier
2014-15 4.33 -
2015-16 4.55
2017-18 4.82 5.15
Overall accuracy (%) (Kappa
statistic)
83.1 (0.74) 76.3 (0.68)
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-3/W6, 2019 ISPRS-GEOGLAM-ISRS Joint Int. Workshop on “Earth Observations for Agricultural Monitoring”, 18–20 February 2019, New Delhi, India
This contribution has been peer-reviewed. https://doi.org/10.5194/isprs-archives-XLII-3-W6-109-2019 | © Authors 2019. CC BY 4.0 License.
113
There is further scope of improving the classification accuracy
through complementary use of cloud free optical data (such as
Sentinel-2) in combination with SAR data.
REFERENCES
Abdikan, S., Sanli, F. B., Ustuner, M., and Calò, F., 2016.Land
Cover Mapping Using Sentinel-1 SARData, Int. Arch.
Photogramm. Remote Sens. Spatial Inf. Sci., XLI-B7, pp. 757-
761, https://doi.org/10.5194/isprs-archives-XLI-B7-757-2016,
2016.
Arfken G. 1985. Lagrange multipliers. Mathematical methods for
physicists. Orlando, FL: Academic Press; p. 945–950. Camps-
Valls G, Bruzzone L. 2005. Kernel-based methods for
hyperspectral image classification. IEEE Trans Geosci Remote
Sens. 43, pp. 1351–1362. doi: http:// dx.doi.org/ 10.1109/ TGRS.
2005.846154.
Bezdek JC. 1981. Pattern recognition with fuzzy objective
function algorithms. Norwell (MA): Kluwer Academic.
Chen, C.F., Huang, S.W., Son, N.T., Chang, L.Y. 2011. Mapping
double-cropped irrigated rice fields in Taiwan using time-series
Satellite Pour I ‘Observation de la Terre data. J Appl Remote
Sens. 5:053528–53528.
Chen, C.F., Son, N.T., Chen, C.R., Chang, L.Y. 2011. Wavelet
filtering of time-series moderate resolution imaging
spectroradiometer data for rice crop mapping using support vector
machines and maximum likelihood classifier. J Appl Remote
Sens. 5:053525–53525.
Dunn, J.C. 1973. A Fuzzy relative of the Isodata process and its
use in detecting compact well-separated clusters. J Cybern. 3, pp.
32–57. doi:http://dx.doi.org/ 10.1080/ 01969727308546046.
Duy Ba Nguyen 1, 2, and Wolfgang Wagner 2 European Rice
Cropland Mapping with Sentinel-1 Data: The Mediterranean
Region Case Study, Water 2017, 9, 392; doi:
10.3390/w9060392.
Galford, G.L., Mustard, J.F., Melillo, J., Gendrin, A., Cerri, C.C.,
Cerri, C.E.P. 2008. Wavelet analysis of MODIS time series to
detect expansion and intensification of row-crop agriculture in
Brazil. Remote Sens Environ. 112, pp. 576–587. Doi: http://
dx.doi.org/10.1016/j.rse.2007.05.017.
Huang, C. Davis, L.S., Townshend, J.R.G. 2002. An assessment
of support vector machines for land cover classification. Int J
Remote Sens. 23, pp. 725–749. doi: http:// dx.doi.org/ 10.1080/
01431160110040323.
Keerthi, S.S., Lin, C.J. 2003. Asymptotic behaviors of support
vector machines with Gaussian Kernel. Neural Comput. 15, pp.
1667–1689. doi:http://dx.doi.org/10.1162/089976603321891855.
Martinez, B., Gilabert, M.A. 2009. Vegetation dynamics from
NDVI time series analysis using the wavelet transform. Remote
Sens Environ. 113, pp. 1823–1842. doi: http:// dx.doi.org/
10.1016/ j.rse.2009.04.016.
McNairn, H., Kross, A., Lapen, D., Caves, R., Shang, J., 2014.
Early season monitoring of corn and soybeans with TerraSAR-
X and RADARSAT-2. Int. J. Appl. Earth Obs. Geoinf. 28, pp.
252–259. http:// dx.doi.org/ 10.1016/ j.jag.2013.12.015.
Nguyen-Thanh Son, Chi-Farn Chen, Cheng-Ru Chen & Vo-
Quang Minh (2018) Assessment of Sentinel-1A data for rice
crop classification using random forests and support vector
machines, Geocarto International, 33:6, pp. 587-601, DOI:
10.1080/10106049.2017.1289555.
Pal, M. 2005. Random forest classifier for remote sensing
classification. Int J Remote Sens. 26, pp. 217–222.
doi:http://dx.doi. org/10.1080/01431160412331269698.
Sakamoto, T., Yokozawa, M., Toritani, H., Shibayama, M.,
Ishitsuka, N., Ohno, H., 2005. A crop phenology detection
method using time-series MODIS data. Remote Sens Environ.
96, pp. 366–374. doi: http:// dx.doi.org/ 10.1016/
j.rse.2005.03.008.
Sesnie, S.E., Finegan, B., Gessler, P.E., Thessler, S, Bendana
ZR, Smith, A.M.S. 2010. The multispectral separability of
Costa Rican rainforest types with support vector machines and
Random Forest decision trees. Int J Remote Sens. 31:2885–
2909. doi: http:// dx.doi.org/ 10.1080/01431160903140803.
Sundaram, R.K. 1996. A first course in optimization theory.
Cambridge (UK): Cambridge University Press.
Whelen, T. and Siqueira, P. 2018. Time-series classification of
Sentinel-1 agricultural data over North Dakota, Remote Sensing
Letters, 9:5, pp. 411-420, DOI:
10.1080/2150704X.2018.1430393.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-3/W6, 2019 ISPRS-GEOGLAM-ISRS Joint Int. Workshop on “Earth Observations for Agricultural Monitoring”, 18–20 February 2019, New Delhi, India
This contribution has been peer-reviewed. https://doi.org/10.5194/isprs-archives-XLII-3-W6-109-2019 | © Authors 2019. CC BY 4.0 License.
114